Pathway-Informed Integration of Expression and Protein Data

Why Bring Pathway Knowledge Into Expression–Protein Integration?

Expression and protein measurements describe different stages of molecular regulation. A gene may show a clear expression change without a corresponding protein change, and a protein signal may appear without a strong expression signal. A pathway knowledge base can help organize such measurements into biologically defined sets, making it possible to ask whether changes in related genes and proteins point in a common direction. It does not, by itself, establish that a pathway is active or that one molecular layer causes a change in another.

The analysis depends on an explicit link between the measured features and the knowledge base. Expression features are usually mapped by gene identifier. Protein features may map to a gene, to several genes, or not at all, depending on the assay and identifier system. Resolve these links before calculating pathway scores, and record the mapping rules. The earlier material on Network Analysis of Cross-Layer Signaling in Cancer concerns cross-layer network structure; here the focus is a pathway-set analysis of expression and protein measurements and a comparison with signals outside the supplied knowledge base.

Definition: A pathway-informed integration analysis uses a specified pathway knowledge base to group mapped expression and protein features, summarizes evidence within each group, and assesses the summaries against an appropriate null distribution. A knowledge-base-supported feature maps to at least one pathway in the chosen knowledge base under the stated mapping rules. An unsupported feature has no such mapping in that knowledge base; this does not mean that it lacks biological relevance.

Set Up the Measurements and Mapping

For a paired study, let $X^{E}_{igc}$ denote normalized log2 expression for subject $i$, gene $g$, and condition $c$, and let $X^{P}_{ipc}$ denote normalized log2 abundance for protein feature $p$. For a specified contrast, such as follow-up minus baseline, define paired changes $d^{E}_{ig}=X^{E}_{ig,\mathrm{followup}}-X^{E}_{ig,\mathrm{baseline}}$ and $d^{P}_{ip}=X^{P}_{ip,\mathrm{followup}}-X^{P}_{ip,\mathrm{baseline}}$. The mean change for each feature estimates its average change in the sampled population under the study design. Positive values indicate increases under this coding; negative values indicate decreases.

A pathway $A$ is represented by a set of mapped genes, together with the protein features assigned to those genes. Decide in advance how to handle multiple protein measurements mapping to the same gene. For example, one may select a protein feature using a documented assay-quality rule or summarize multiple features using a prespecified aggregation. Counting the same gene multiple times because it has several protein identifiers can give that gene unintended weight. Exclude unmapped features from pathway scores but retain them for the comparison with unsupported signals.

If pathway members have different measurement scales or precision, a common approach is to calculate signed feature-level statistics, such as standardized statistics from a prespecified differential analysis. Denote these by $z^E_g$ and $z^P_g$, where positive and negative signs follow the chosen contrast. They are evidence summaries, not direct measurements of pathway activity. Specify the feature-level model and its assumptions before using its statistics; for paired differences, a mean-based model requires independent subjects and, for exact small-sample $t$ inference, normally distributed differences. The pathway permutation approach below has its own exchangeability assumptions and does not make those feature-level assumptions disappear if the $z$ values are treated as fixed inputs.

Definition: For pathway $A$, let $G_A$ be its mapped genes and let $w_g$ be prespecified gene weights satisfying $w_g\geq 0$ and $\sum_{g\in G_A}w_g>0$. Define the expression and protein pathway summaries by $$ T_A^E=\frac{\sum_{g\in G_A}w_g z_g^E}{\sum_{g\in G_A}w_g}, \qquad T_A^P=\frac{\sum_{g\in G_A}w_g z_g^P}{\sum_{g\in G_A}w_g}. $$ A positive summary indicates a positive weighted average direction under the chosen coding; it does not imply that every pathway member has a positive statistic. The nonnegative weights and positive total weight are necessary for this directional interpretation.

What a Pathway Score Does—and Does Not—Show

The two summaries should first be inspected separately. If both are positive, the average evidence in both layers points toward an increase; if both are negative, it points toward a decrease. Opposite signs indicate discordance in the average directions. These patterns are descriptive until calibrated against a null distribution. A combined score can be defined in advance, but a numerical combination should not be treated as having a known reference distribution unless its assumptions justify one.

For example, a prespecified directional-concordance statistic can be $J_A=\min(T_A^E,T_A^P)$ when the scientific question is whether both layers increase. A large value requires both summaries to be large and positive. For a question about concordant changes in either direction, define the direction using an independent rule or test the positive and negative alternatives separately with suitable error control. Choosing whichever direction looks strongest after seeing the data changes the null distribution and must be accounted for.

Pathway members are often correlated. The effect of this dependence can be seen by writing a weighted mean as $T=w^\mathsf{T}z/\sum_g w_g$, where $z$ is the vector of member statistics. If the null covariance matrix of those statistics is $\Sigma$, then $$ \operatorname{Var}(T)=\frac{w^\mathsf{T}\Sigma w}{(\sum_g w_g)^2}. $$ This follows by applying $\operatorname{Var}(a^\mathsf{T}z)=a^\mathsf{T}\Sigma a$ with $a=w/\sum_g w_g$. The diagonal terms describe individual variances; off-diagonal terms describe dependence between members. Ignoring positive covariance can underestimate variability and make a pathway look more significant than it is.

A practical calibration is to permute condition labels at the subject level and recompute the full pathway statistic. In a paired design, swap the two condition labels within each subject, applying the same swap to every expression and protein feature for that subject. This preserves the observed cross-feature and cross-layer dependence. Such permutations provide exact randomization inference when the condition assignments are exchangeable under the null, as in an appropriate randomized paired design. For observational paired data, sign-flip inference instead requires a suitable symmetry or exchangeability assumption for the paired differences; it is not automatically valid merely because measurements are paired.

Freeze the analysis definitions.
Specify the contrast, normalization, identifier mapping, duplicate-feature rule, pathway collection, weights, and direction of the test before interpreting results.

Calculate layer-specific evidence.
Estimate or summarize expression and protein changes using the stated feature-level procedures, preserving the sign convention in both layers.

Summarize and calibrate.
Calculate pathway scores, then use a null procedure that respects the design and the dependence among features. Apply multiplicity control across the pathways tested.

Compare supported and unsupported signals.
Report results for mapped pathways alongside summaries for eligible unmapped features, keeping their different interpretation and coverage explicit.

Worked Example: Checking Paired Changes in Both Layers

Suppose four hypothetical subjects have normalized log2 measurements at baseline and follow-up. For one mapped gene, the expression changes are $0.8, 1.0, 0.6,$ and $1.2$ log2 units. For its corresponding protein feature, the changes are $0.4, 0.8, 0.6,$ and $0.6$ log2 units. The paired design and common contrast make the signs comparable, although expression and protein remain distinct measurements.

For expression, the mean paired change is $$ \bar{d}^{E}_g=\frac{0.8+1.0+0.6+1.2}{4} =\frac{3.6}{4}=0.9000\text{ log2 units}. $$ The deviations from the mean are $-0.1, 0.1, -0.3,$ and $0.3$. Thus $$ s_E=\sqrt{\frac{0.01+0.01+0.09+0.09}{4-1}} =\sqrt{0.06667}=0.2582\text{ log2 units}, \qquad SE_E=\frac{0.2582}{\sqrt{4}}=0.1291\text{ log2 units}. $$

For protein, the mean change is $$ \bar{d}^{P}_g=\frac{0.4+0.8+0.6+0.6}{4} =\frac{2.4}{4}=0.6000\text{ log2 units}. $$ The deviations are $-0.2, 0.2, 0,$ and $0$, giving $$ s_P=\sqrt{\frac{0.04+0.04}{3}}=0.1633\text{ log2 units}, \qquad SE_P=\frac{0.1633}{\sqrt{4}}=0.08165\text{ log2 units}. $$

Both estimated changes are positive. On the log2 scale, these correspond to geometric fold changes of $2^{0.9}\approx1.866$ for expression and $2^{0.6}\approx1.516$ for protein. These estimates describe this hypothetical paired sample; they do not show that the gene or its pathway caused a response. Small-sample uncertainty also cautions against treating directional agreement alone as strong evidence.

Worked Example: Calculating Directional Pathway Summaries

Consider a hypothetical pathway with three mapped genes, equal for expression and protein mapping, and prespecified weights $w=(1,1,0.5)$. Suppose the feature-level signed statistics are $z^E=(3.4,2.8,-0.5)$ and $z^P=(2.1,1.7,-1.0)$. The weights are nonnegative and sum to $2.5$, so the summaries retain an average-direction interpretation.

For expression, $$ T_A^E=\frac{1(3.4)+1(2.8)+0.5(-0.5)}{1+1+0.5} =\frac{3.4+2.8-0.25}{2.5} =\frac{5.95}{2.5}=2.380. $$ For protein, $$ T_A^P=\frac{1(2.1)+1(1.7)+0.5(-1.0)}{2.5} =\frac{2.1+1.7-0.5}{2.5} =\frac{3.3}{2.5}=1.320. $$

Both pathway summaries are positive, and the concordance score for the prespecified “increase in both layers” question is $J_A=\min(2.380,1.320)=1.320$. The third member is negative in each layer, so the positive pathway averages do not mean every member increased. Nor does $J_A=1.320$ have a p-value without a valid reference distribution that accounts for the analysis design and pathway-member dependence.

Assessing Evidence and Comparing Knowledge-Base Coverage

A pathway’s empirical p-value can be calculated from $B$ valid null permutations. For a prespecified upper-tail statistic, if $b$ permutation statistics are at least as large as the observed statistic, the commonly used Monte Carlo estimate is $(b+1)/(B+1)$; the added one prevents a zero estimate from a finite set of random permutations. With exhaustive permutations, calculate the exact tail proportion over all permitted assignments. When many pathways are tested, adjust for multiple comparisons, for example by controlling the false discovery rate with the Benjamini–Hochberg procedure. Report the number of tested pathways, the correction method, and whether the tested collection was selected before or after examining the data.

The comparison with unsupported signals is part of interpretation, not a contest in which mapped findings automatically win. Use a stated feature universe: for example, all reliably measured genes and proteins that passed the same assay-quality and differential-analysis criteria. Summarize unsupported features using the same contrast and, where appropriate, comparable statistics. A pathway score supported by annotations is interpretable relative to that knowledge base; an unmapped signal may reflect incomplete annotation, assay-specific measurement, identifier failures, or biology not represented in the resource. Conversely, a coherent-looking unsupported signal is not evidence of a newly established pathway without additional validation.

Approach	What it contributes	Main caution
Separate feature-level analysis	Shows which expression and protein features change and in which direction.	Does not itself summarize coordinated evidence by pathway.
Pathway-set integration	Summarizes mapped features within defined biological sets and compares layers.	Depends on mapping, pathway definitions, and dependence-aware calibration.
Over-representation analysis	Tests whether a selected feature list contains more pathway members than expected under a defined background.	Depends on the selection threshold and correct background; it can discard effect size and direction.
Unannotated-feature summary	Shows whether notable signals fall outside the supplied pathway collection.	Absence of annotation is not evidence of absence of biological function.

Worked Example: Exact Paired Permutation and an Unsupported Signal

In a hypothetical study with four paired subjects, suppose a prespecified pathway-level composite change is $D=(1,1,1,1)$ standardized units, and the observed statistic is the mean $\bar{D}=1$. The analysis tests the one-sided alternative that the pathway composite increases. Assume paired sign flips are valid under the null. There are $2^4=16$ possible sign assignments, and the permuted mean can be no larger than 1. Only the assignment with all four signs positive reaches 1, so $$ p=\frac{\#\{\bar{D}^{*}\geq 1\}}{16}=\frac{1}{16}=0.0625. $$ This does not meet a 0.05 threshold, even before adjusting for testing multiple pathways.

Now suppose four equally weighted, eligible features without a pathway mapping have hypothetical paired composite changes $2,-2,2,-2$ standardized units. Their observed mean is $$ \bar{D}_{\mathrm{unsupported}}=\frac{2-2+2-2}{4}=0. $$ A two-sided test of this mean has a permutation p-value of 1, because every sign-flipped mean has absolute value at least 0. The supported pathway result and this unsupported summary are not directly interchangeable biological claims: they have different feature membership and interpretation. Together they show why reporting coverage and the exact tested feature universe matters. A failure to find a directional unsupported summary does not prove that all unmapped features are unimportant.

Diagnostics, Limitations, and Common Pitfalls

Before interpreting a pathway result, check that the mapping rate is reported separately for each assay, identifiers resolve consistently, and duplicate mappings do not inflate a pathway’s score. Inspect the direction and magnitude of both layer summaries, rather than reporting only a combined value. Check whether results are driven by a single high-scoring feature, whether alternative reasonable pathway versions change the conclusion, and whether the permutation procedure matches the design. A pathway with very few measured members or poor coverage may not be comparable to a well-measured pathway. These checks diagnose fragility; they do not establish that a pathway is biologically active.

A common mistake is to reason, “The pathway average is positive, therefore every gene and protein in the pathway increased.” That is false: a weighted average can be positive while some member statistics are negative, as the preceding example shows. The correction is to report the average direction and the member-level pattern separately. Another mistake is to treat every feature as independent when calculating uncertainty; correlated members invalidate that shortcut. Use a calibration that respects dependence, or state and justify a model for it.

Pathway results are conditional on the chosen database, its version, the mapping policy, and the measured feature universe. Different resources can define pathway membership differently, while well-studied biology may be more densely annotated than less-studied biology. Pathway-level evidence is therefore not a causal claim, a direct assay of pathway activity, or proof that omitted biology is irrelevant. Report enough detail for a reader to reconstruct the mapping, score, null procedure, and multiplicity adjustment.

Key takeaway: Pathway-informed integration organizes expression and protein evidence using an explicit knowledge base. Credible interpretation requires transparent mapping, separate layer summaries, dependence-aware inference, multiplicity control, and an honest account of signals the knowledge base does not cover.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If a protein feature maps to two genes, what should be decided before calculating pathway scores?
A pathway has nonnegative weights summing to 4 and a weighted expression-statistic numerator of 6. What is its expression summary, and what does its sign mean?
Why can treating pathway-member statistics as independent produce misleading uncertainty?
What does a feature’s lack of a pathway mapping establish, and what does it not establish?

Solutions. (1) Specify and report a mapping and duplicate-handling rule, such as a prespecified protein-to-gene assignment or aggregation, so the feature is not unintentionally counted multiple times. (2) The summary is $6/4=1.5$, a positive weighted average direction under the chosen coding; it does not imply every member is positive. (3) Correlated members have nonzero covariance, which contributes to the variance of a weighted sum; ignoring those terms can underestimate uncertainty. (4) It establishes only that the feature is unsupported by the specified knowledge base under the stated mapping rules. It does not establish that the feature has no biological relevance.